Knowledge-Based Extrapolation of Neural Network Model for Transistor Modeling

Abstract

Artificial neural network (ANN) is a useful technique for active device modeling. However, it shows limitations in the extrapolation region. To address this issue, we propose a novel knowledge-based neural network (KBNN) method. The KBNN technique consists of three submodels and their transition mechanisms. One submodel is a pure ANN model which is used for training data region. Two additional submodels are used for the extrapolation region. The proposed method ensures that the output and derivatives of ANN and extrapolation models match at the boundary of the measurement data. This keeps the KBNN model smooth and consistent, making it suitable for transistor design over a broad range. The precision, smoothness, and consistency of the proposed method are verified with a $2\times 250~\mu $ m GaN HEMT device modeling. The results show that the KBNN model provides physically reasonable predictions over a wide extrapolation region.